Invisible Steganography via Generative Adversarial Network
share
Steganography and steganalysis are main content of information hiding, they always make constant progress in confrontation. There is large consent that measures based on deep learning have outperformed conventional approaches in steganalysis, which have shown that deep learning is very promising for the information hiding area. And in the last two years, there are also several works using deep learning to achieve the procedure of steganography. While these works still have problems in capacity, invisibility and security. We proposed a new steganography model based on generative adversarial network named as ISGAN. Our model can conceal a secret gray image into a color cover image, and can reveal the secret image successfully. To improve the invisibility, we select a new concealing position and get excellent results. And with the help of GAN's adversarial training, our model can improve the security. In addition, we use a new loss function which is more appropriate to steganography, can boost the training speed and generate better stego images and revealed secret images. Experiment results show that ISGAN can achieve start-of-art performance on LFW, Pascal VOC2012 and ImageNet datasets.
READ FULL TEXT